A novel hybrid artificial neural network - Parametric scheme for postprocessing medium-range precipitation forecasts

Highlights

• A novel, hybrid artificial neural network – EMOS postprocessing scheme is proposed.

• The hybrid scheme allows for simultaneous postprocessing of precipitation forecasts for multiple seasons and lead times.

• The scheme outperforms existing EMOS schemes as judged by the skills of postprocessed probabilistic precipitation forecasts.

Abstract

Many present-day statistical schemes for postprocessing weather forecasts, in particular precipitation forecasts, rely on calibration using prescribed statistical models to relate forecast statistics to distributional parameters. The efficacy of such schemes is often constrained not only by prescribed predictor-predictand relation, but also by arbitrary choices of temporal window and lead time range for training. To address this limitation, we propose an end-to-end, computationally efficient hybrid postprocessing scheme capable of producing full predictive distributions of precipitation accumulation without explicit stratification of forecast-observation pairs by forecast lead time and season. The proposed framework uses the censored, shifted gamma distribution (CSGD) as the predictive distribution but uses an artificial neural network (ANN) to estimate the distributional parameters of CSGD through a unified approach. This approach, referred to as ANN-CSGD, allows for simultaneous estimation of distributional parameters over multiple lead times and seasons in a single model by incorporating the latter variables as predictors to the ANN. We test our proposed ANN-CSGD model for postprocessing of ensemble mean forecasts of 24-h precipitation totals over selected river basins in California, at one- to seven-day lead times, from the Global Ensemble Forecast System (GEFS). The probabilistic quantitative precipitation forecasts (PQPFs) from the ANN-CSGD, are more skillful overall than those from the benchmark CSGD and the Mixed-type meta-Gaussian distribution (MMGD) models. The ANN-CSGD PQPFs highly improve the performance of those from CSGD in predicting the probability of precipitation (PoP) and are also much sharper and reliable at higher precipitation thresholds. We demonstrate how the hybrid approach, by using the entire available training data and its modified formulation, efficiently represents interactions between GEFS forecasts and season/lead times, thus leading to enhanced predictive performance.

Introduction

Statistical postprocessing techniques are increasingly used to improve the reliability and skill of real time probabilistic quantitative precipitation forecasts (PQPFs) produced by numerical weather prediction (NWP) models. Broadly speaking, these techniques can be categorized as nonparametric and parametric ones. A prominent example of the former is the Analog approach (Hamill and Whitaker, 2006; Hamill et al., 2015). The parametric techniques rely on prescribed parametric forms of conditional (predictive), joint and marginal distributions, and employ various techniques ranging from regression to the method of moments, and their variants, for estimating distributional parameters. Many of the modern parametric approaches fall under the broad umbrella of Ensemble Model Output Statistics (EMOS; Gneiting et al., 2005), also known as nonhomogeneous regression. As the name implies, the EMOS approaches use prescribed predictive distributions and relate distributional parameters to ensemble statistics through a set of regression equations (Scheuerer and Hamill, 2015; Zhang et al., 2017; Stauffer et al., 2017).

The extent to which postprocessing techniques have improved forecast skill has varied in practice (Li et al., 2017; Wilks, 2018; Vannitsem et al., 2020). There are several common limitations in postprocessing methods adopted to date. Among the frequently cited are the inflexible and subjective way of selecting predictors, structural rigidity that makes it difficult to integrate ancillary predictors, and the ad hoc way of determining spatial-temporal training domains (see related discussions in Rasp and Lerch (2018)). The advent of machine learning techniques offers many new opportunities to address these limitations. Relative to the parametric approaches, EMOS techniques included, some of the recent machine learning techniques offer flexibility in identifying predictors, in integrating ancillary information, and in capturing complex, nonlinear predictor-predictand relationships that are difficult to characterize parametrically (see, e.g., Taillardat et al., 2019). Particularly promising are the various artificial neural networks (ANNs) which have been known for their ability to model nonlinear dependencies. Recent years have seen an explosion of ANN-based prediction paradigms (Liu et al., 2016; Brenowitz and Bretherton, 2018; Gentine et al., 2018; Rasp et al., 2018; Chapman et al., 2019; Cloud et al., 2019; Gagne et al., 2019; Lagerquist et al., 2019). Yet, the use of these techniques in the context of postprocessing remains relatively limited. Rasp and Lerch (2018) is perhaps the first attempt of this nature. The authors explored a hybrid scheme that retains a parametric form of the predictive distribution of 2-m temperature but relies on ANNs to estimate the distribution parameters from the ensemble statistics of 2-m temperature as well as ancillary variables. Scheuerer et al. (2020), in a similar vein, developed an ANN-based scheme for producing 7-day accumulated PQPFs at subseasonal range (2‒4 weeks) from NWP ensemble forecasts, and showed that the PQPFs thus generated broadly outperforms climatology. Other studies of note include Bremnes (2020) wherein ANN was used for postprocessing wind speed forecasts. Collectively, these studies indicate that embedding local information and incorporating ancillary forecast variables can lead to clear, discernible improvements in forecast skills. They further suggest that ANN models, contrary to the common perception of being black boxes, can help uncover, and offer physical insights to the meteorological processes that underpin the links between predictors and predictands.

Inspired by the successes of recent ANN-based postprocessing approaches, and motivated by the broader need for improving the skill of PQPF while circumventing limitations inherent in existing EMOS schemes, we propose a hybrid ANN-nonhomogeneous regression-based scheme capable of postprocessing precipitation forecasts at multiple lead times and seasons in a unified way. The proposed scheme retains the parametric form of the predictive distribution of precipitation proposed by Scheuerer and Hamill (2015) and Baran and Nemoda (2016), but departs from the conventional EMOS by using ANNs to relate NWP forecasts to the distributional parameters. The potential advantages of the proposed scheme, which we will henceforth refer to as ANN-CSGD are three-fold. First, this scheme does not require an explicit prescription of predictor-predictand relationships as is currently done in EMOS models - it can discover and integrate arbitrary nonlinear relationships through training. Second, the training of the model can be done using the entire data archive and thereby obviate the need for explicit treatment of lead time-based and seasonally varying NWP forecast errors. Third, it can account for seasonal variations in the interaction between NWP forecasts and temporal predictors.

In this paper we describe and evaluate the proposed scheme which relies only on the ensemble mean of NWP forecasts as the major predictor. The evaluation is conducted for sub-basins within three selected river basins in California. The proposed scheme is applied to postprocess Global Ensemble Forecast System (GEFS; Hamill et al., 2013) precipitation reforecasts along with two benchmark schemes. The first is the single predictor version of the censored, shifted gamma distribution (CSGD; Scheuerer and Hamill, 2015). The second is the Mixed-type Mata-Gaussian Distribution (MMGD; Wu et al., 2011), which has been the standard method in the U.S. National Weather Service (NWS) Hydrologic Ensemble Forecast Service (HEFS; Demargne et al., 2014). Our overarching hypothesis is that the flexibility accorded by the ANN-based model in establishing complex predictor-distributional parameter relationships, in determining temporal training windows, and in lumping forecasts for different lead times, will help the proposed scheme attain superior predictive performance relative to the benchmarks.

The reminder of this paper is organized as follows. Section 2 describes the proposed ANN-CSGD scheme as well as the benchmark methods, data, and experimental setup. Section 3 presents the outcomes of the experiments and section 4 summarizes the findings and discusses future possible extensions.